SelfDefend introduces a novel framework enabling LLMs to detect and counter jailbreak attempts without external tools or significant latency.

• Handles multiple attack types including human-based, optimization-based, and indirect jailbreaks
• Operates with negligible processing delays (5-50ms overhead)
• Compatible with both closed and open-source LLM deployments
• Demonstrates strong practical effectiveness against evolving attack methods

This research addresses critical security challenges in AI deployment, providing a solution that balances robust protection with operational efficiency—essential for organizations implementing LLMs in production environments.

SelfDefend: LLMs Can Defend Themselves against Jailbreaking in a Practical Manner